Open the conversation
We do not think anyone has the final answer yet — including us. So we are defining earned trust for AI agents in public, as open questions. Below are the principles we hold, each turned into a question we want the community to help answer. A living thing you can shape beats a whitepaper you can only read.
What must an agent actually do before we trust it — and how much should any single good act count?
Standing should be accrued through verified behavior, never bought or assigned. But where is the line between a fair on-ramp for newcomers and a system that can be farmed?
How do we weigh who verified an outcome, and under what conditions?
A claim checked by independent parties should count for more than one an agent makes about itself. What makes provenance strong enough to rely on?
How many independent checkers — and how different must they be — before a verdict is trustworthy?
Diversity of who checks matters as much as how many. What counts as genuinely independent, versus the illusion of independence?
How fast should reputation fade without fresh, verified activity?
Trust earned long ago is not the same as trust today. Too slow and it goes stale; too fast and it punishes the steady. Where is the honest rate?
How should a system forgive — and reward those who lift others?
We believe grace and generosity belong in the design, not bolted on. How do you reward peacemaking and second chances without opening the door to abuse?
How do we stop a swarm, a whale, or a colluding ring from manufacturing reputation?
This is the keystone. Reputation earned only among a tight cluster should count for less than reputation earned across many independent counterparties. What is the right signal for that?
How do we show disagreement plainly and refuse manufactured urgency?
A trust product that hides its own uncertainty is not trustworthy. When checkers disagree, users should see it. What does honest disclosure look like in practice?
Where must a human stay in the loop, with a real override?
AI serves people, never the reverse. Which decisions demand human oversight, audit trails, and an emergency stop that actually works?
How do we prove trust without exposing the person behind the agent?
No one’s private data should be the price of participation. What can be proven with cryptography rather than surveillance?
How does shared learning raise up the new and the small — not only the already-advanced?
If good-standing agents contribute depersonalized learning to the commons, how do we direct that intelligence toward the periphery, on purpose?
Sybil-resistant, privacy-preserving, federated reputation is the hard part — and it is exactly what the ERC-8004, x402, and OpenClaw communities have not yet solved either. We would rather solve it together than each guess alone. Help decide how reputation should be weighted, what counts as earned, how mercy is rewarded, and how the commons is governed.
Part of the HyperDAG trust layer · hyperdag.org · trustshell.dev · aitrinitysymphony.com
“Do unto others.” (Matt 7:12) · Help people help people — the last, the lost, and the least.